pi-subagent

Minimal subagent runtime for Pi.

pi-subagent adds one focused tool: subagent. It gives Pi the essentials for isolated worker runs — parallel fan-out, sandbox/worktree controls, durable artifacts, and async status.

It is intentionally small, so you can add it to a project when you need subagents and remove it when you do not.

npm package: @agwab/pi-subagent

Installation

pi install npm:@agwab/pi-subagent

Then reload Pi.

Requires Node.js >=22.19.0 on macOS or Linux. Native Windows is not supported (POSIX process groups, tmux, and which-based Pi discovery); use WSL2.

For local development, add this package as a Pi extension source and reload Pi.

Quick usage

Use it when you want Pi to spin up a separate worker instead of doing everything in the parent session:

Run three reviewers in parallel for this change.

Run this check in a sandboxed worker and report the artifact paths.

Start a background audit and let me inspect it in /subagent panel.

What it does

Tool: subagent

Sandbox

Run workers in an isolated local execution boundary.

{
  "sandbox": true,
  "agent": "checker",
  "task": "Run a local check and report the artifact paths."
}

sandbox: true denies all network access. Model-backed sandboxed runs must allow their provider endpoint explicitly:

{
  "sandbox": { "allowedDomains": ["api.anthropic.com"] },
  "agent": "implementer",
  "task": "Make the requested local change and run the checks."
}

Worktree

Isolate parallel or mutating tasks in managed git worktrees. Workspaces default to shared; request worktree: true explicitly for tasks that mutate files in parallel.

{
  "worktree": true,
  "agent": "implementer",
  "task": "Make the requested local change in an isolated worktree."
}

Agent

Inject Pi subagent markdown definitions from global or project agent directories.

{
  "agent": "reviewer-security",
  "task": "Review the current diff for security risks."
}

Agent markdown can live in ~/.pi/agent/agents/*.md or .pi/agents/*.md. Agent-level tools declarations are an authority ceiling; call-level tools can narrow them but not expand them. A systemPrompt override replaces the agent prompt body, not the agent's frontmatter policy.

Type

Use one structured schema for single, parallel, async, and existing-run calls. action defaults to run. Each execution is a run; each launch is an attempt.

Single:

{
  "agent": "reviewer",
  "task": "Review the current diff and summarize the highest-risk issues."
}

Parallel launches independent runs concurrently:

{
  "tasks": [
    { "agent": "reviewer-security", "task": "Review the current diff for security risks." },
    { "agent": "reviewer-performance", "task": "Review the current diff for performance risks." },
    { "agent": "reviewer-test-coverage", "task": "Review the current diff for missing tests." }
  ]
}

Existing run:

{ "action": "status", "runId": "run_..." }

Recent runs can be addressed by runId even when they were launched from another cwd; legacy records still resolve from the explicit or current cwd.

Panel

Inspect runs, attempts, artifacts, and log tails in a live TUI. The panel defaults to the current Pi session, can switch to current cwd or all indexed runs, and includes status filters plus a scrollable detail pane. It shows active and recent terminal runs by default, with in-panel m to show more, and counts stale/malformed run pointers without exposing raw session ids.

Open the run monitor:

/subagent panel

Code API

Orchestrators can use the same runtime directly:

import { runSubagent, getSubagentStatus } from "@agwab/pi-subagent/api";

const run = await runSubagent({ agent: "reviewer", task: "Review this diff.", async: true });
const status = await getSubagentStatus({ runId: run.runId });

Detailed docs

• docs/usage.md — full argument reference, code API, action behavior, backend selection, sandbox/worktree behavior, artifacts, and validation notes.